AdaBoost.RDT: AdaBoost Integrated With Residual-Based Decision Tree for Demand Prediction of Bike Sharing Systems Under Extreme Demands

Boosting algorithms are widely used for predicting demand in bike-sharing systems (BSSs). However, these systems often encounter sudden spikes in demand (extreme demand). Ordinary boosting algorithms tend to be biased toward extreme demands, leading to increased prediction errors in other scenarios. Noise-robust boosting algorithms perform well with normal samples; however, for normal samples in datasets containing extreme demands, their accuracy remains poor for extreme demand samples. To address these limitations, we propose a novel boosting algorithm, AdaBoost.RDT, which integrates adaptive boosting with a residual-based decision tree. Our approach aims to enhance prediction accuracy for extreme demand scenarios without compromising performance in normal situations. By incorporating a decision tree (DT) model at each boosting iteration to predict residuals from the base model, we effectively identify and improve predictions for underestimated extreme demands. AdaBoost.RDT was compared with six boosting algorithms, including noise-robust variants, using Seoul Bike and Daejeon Bike data. Experimental results demonstrated that the DT model within AdaBoost.RDT effectively distinguished between over- and under-estimated samples, significantly reducing prediction errors for extreme demand scenarios with compromised accuracy for very low demands. On the stance in operating a shared bicycle service, it is important to alleviate the customer dissatisfaction caused by not being able to rent bicycle encouraged by extreme events. Therefore, it should be achieved even if it requires compromised accuracy for very low demands.